Qwen3.8 27B

User reports Qwen3.8 27B at 54.68 t/s on an RTX 5070 Ti 16GB, with about 55 t/s overall. Setup is llama.cpp with a 4.2BPW GGUF (11.8GB) and q8_0/q4_0 KV cache at 110000 context, with the display driven by the 9800X3D iGPU to free the card's VRAM. The prompt eval figure of 56.71 t/s covers only 4 tokens, so it is not recorded as prefill. The user also has a GPU overclock of +350 core, +4000 memory and a 350W BIOS, and 32GB DDR5.